Pollen Grain Classification Based on Ensemble Transfer Learning on the Cretan Pollen Dataset.

Area of Science:

• Botany
• Computer Science
• Food Science

Background:

• Pollen identification is crucial for honey's botanical certification.
• Manual microscopic analysis (melissopalynology) is labor-intensive and prone to errors.
• Automated methods are needed to improve efficiency and consistency.

Purpose of the Study:

• To evaluate deep learning models for classifying 20 pollen types using the Cretan Pollen Dataset.
• To assess the performance of transfer and ensemble learning techniques for pollen classification.
• To investigate the effectiveness of a trained model on real honey samples.

Main Methods:

• Utilized transfer learning and ensemble methods on the Cretan Pollen Dataset.
• Trained deep learning models for pollen grain image classification.
• Performed a preliminary case study applying the best model to honey-based pollen images.

Main Results:

• Achieved high performance on the Cretan Pollen Dataset: 97.5% accuracy, 96.9% sensitivity, 97% precision, 96.89% F1 score, and 0.9995 AUC.
• The best-performing model showed poor results on honey-based pollen images (AUC of 0.52).

Conclusions:

• Deep learning models, particularly with transfer and ensemble learning, are highly effective for classifying pollen types in controlled datasets.
• Significant fine-tuning on honey-specific pollen data is necessary for practical application in honey analysis.
• Further research should focus on adapting models for real-world melissopalynology challenges.